Multivariate filter methods for feature selection with the γ -metric.

Area of Science:

• Machine Learning
• Biomedical Informatics
• Statistical Analysis

Background:

• The gamma-metric is commonly used for feature importance scoring in classification.
• Existing methods lack specific search directions, limiting their optimization.
• This study addresses this by integrating the gamma-metric with defined search strategies.

Purpose of the Study:

• To develop and evaluate a new methodology for multivariate feature selection in classification.
• To associate the gamma-metric with specific search directions, creating distinct methods.
• To compare these novel methods against conventional approaches.

Main Methods:

• A simulation study was conducted to assess the performance of the new methodology.
• The methods were compared based on classification accuracy, feature selection stability, and computational time.
• The methodology was applied to a real-world task: detecting atrial fibrillation.

Main Results:

• The proposed methods successfully identified informative features and maintained predictive performance in simulations and AF detection.
• However, gamma-metric based methods showed reduced efficiency in excluding non-informative features with highly correlated data and large datasets.
• The forward search direction combined with the gamma-metric proved particularly effective.

Conclusions:

• A combination of forward search and the gamma-metric offers a robust approach for feature selection.
• The backward search direction may lead to local optima, suggesting areas for algorithmic improvement.
• The developed methods provide a valuable tool for feature selection in complex classification problems.